Implementation
How to clean up HubSpot before you switch on the AI agents – a 30-day plan
AI agents on messy data just make mistakes faster. A week-by-week plan for duplicates, fields, definitions and owners in HubSpot, with a checklist.
By Viktor Olsson, founder and RevOps consultant
Implementation · AI· Updated · 20 min read
You clean up HubSpot for the AI agents in four weeks, one area at a time. In week 1 you merge the duplicates, companies first. In week 2 you prune the fields and make the important ones required at the deal stage where the answer is known. In week 3 you write down what each lifecycle stage and deal stage means, and sales, marketing and customer success sign it off. In week 4 you give every customer an owner, connect everyone's inbox and name one person who owns data quality. Then you switch on one agent at a time.
The plan is written for B2B companies, from teams of 3 people to organisations with 30 sales reps. Why the data decides what the AI can do is covered in Context is king and AI in HubSpot. This article is about how to do it.
Why agents on messy data make mistakes faster
A rep who finds the same customer twice in search asks a colleague which one is right. An agent asks nobody. It picks one of the records, does its job and moves on to the next customer. That's the whole point of agents: they do the same thing a hundred times without anyone watching. That goes for the mistakes too.
Take the example customer Lövdala Bygg AB. The rep creates the company with the customer's domain when the first deal starts. A connection from the accounting system, Fortnox, then creates Lövdala Bygg Aktiebolag when the first invoice goes out. Companies created by an integration aren't automatically deduplicated by domain. The customer's CFO ends up as a contact on the second record. When she emails support about an incorrect invoice, the ticket is associated there, and the rep never sees it.
The next day the rep is about to call the customer about the renewal and asks Breeze Assistant to summarise the customer first. The assistant reads the record with the deals and replies that there are no open tickets. The answer looks complete, and the quote goes out while the invoice sits unresolved in support.
The AI does nothing wrong. It reads what is on the record it finds. The mistake is in the data, and it doesn't show in the answer.
A worked example: what duplicates and empty fields cost
Duplicates and empty fields already cost time today, before any agent is switched on. Here's a worked example for a sales team. Swap the numbers for your own.
| Assumption | Value |
|---|---|
| Reps and customer success managers (CSMs) working in HubSpot | 12 |
| Times a week each of them finds a duplicate or an empty field and has to sort it out | 3 |
| Time each time: find the right record, ask a colleague, fix it | 10 minutes |
| Working weeks a year | 44 |
| Hours a year | 264 |
| Salary cost per hour including employer contributions | SEK 500 |
| Cost per year | SEK 132,000 |
With agents the cost takes a different shape. Assume an agent writes 40 summaries a week ahead of calls and meetings, and that one customer in ten exists twice. Then four summaries a week miss part of the customer, 176 in a year. Nobody notices them, because they look just as complete as the others.
Day 0: measure where you are today
Write down five numbers before you start. Pull the same numbers again after 30 days, and then every month.
| Measure | Where in HubSpot |
|---|---|
| Number of duplicate pairs for companies and for contacts | Manage duplicates, from the companies and contacts lists |
| Fields with no data, unused fields and fields that look alike | The data quality overview, Property insights tab |
| Formatting issues | The data quality overview, Formatting issues tab |
| Companies and open deals without an owner | Filter: owner is unknown |
| Open deals without an amount or with a close date in the past | Filter in the deals list |
Also check what your licence includes. The data quality overview is available from Starter and is in beta. The duplicates tool needs Professional or Enterprise.
The plan week by week
-
Week 1
Duplicates
Merge companies first, then contacts. Fix the source that creates them.
Done when the pairs are reviewed and the source is fixed.
-
Week 2
Fields
Archive what isn't used, turn free text into dropdowns and require the right fields at the right deal stage.
Done when every field on the record has a reason to be there.
-
Week 3
Definitions
Write down lifecycle stages and deal stages. Sales, marketing and customer success sign them off.
Done when the page is signed off and the stages in HubSpot use the same names.
-
Week 4
Owners and context
Owners on every record, connected inboxes, Context Home and one person who owns data quality.
Done when no record lacks an owner and the monthly review is booked.
Take the weeks in order. Fields are easier to prune once the duplicates are gone, and definitions are easier to write once there are fewer fields.
Week 1: merge the duplicates
Where in HubSpot. Open the companies or contacts list and choose Actions → Manage duplicates. HubSpot compares field values as new records are created and suggests pairs. For contacts it compares fields such as name, email, phone and company name. For companies it compares fields such as domain, name, country and phone.
Companies first. Contacts, deals and tickets hang off companies. If you merge the companies first, everything that belongs to the customer ends up on one record. Duplicate contacts are then easier to spot, because they land on the same company.
Which record wins. The record you merge into becomes the primary one. Its values are kept, and where a field is empty it's filled in from the other record. Activities and associations from both records end up on the new one. For contacts, the other email address becomes an additional address, and the lifecycle stage furthest down the funnel is kept. Our rules for which record should win:
- Companies: the record with the right domain and with the deals and tickets. If that's unclear, the oldest one.
- Contacts: the record with the email address the person uses today, usually their work address, and the most recent activity.
With Data Hub Professional or Enterprise you can merge many pairs at once. You then choose what decides the primary record, for example the most recent activity or the record created first. Since summer 2026 there are also custom rules for which record wins, for example the one with the most fields filled in. They're in public beta.
A merge can't be undone. HubSpot can't split a merged record again. So do this:
- Filter the pairs by how similar they are. HubSpot shows a similarity score, and you take the closest matches first.
- Add columns for the fields you decide on, for example domain, owner and number of deals.
- Look at the merge preview for the first twenty pairs before you merge any more.
- Reject pairs that aren't duplicates. A rejected pair can be restored within 14 days.
Fix the source. Duplicates usually come from imports, from forms where people enter their personal address, and from integrations. Companies created through the API, including by installed apps, aren't deduplicated by domain. If many duplicates come from the same integration, fix it first. Otherwise they'll be back next month. We describe how HubSpot and the accounting system can be connected in From quote to invoice in HubSpot. If your customers include chains and franchises, new companies can also be linked to the right parent company automatically, see Link child companies to the right parent company in HubSpot.
For Swedish B2B customers, the organisation number (organisationsnummer) is a good safeguard. Make it its own field that requires unique values, which you set when the field is created. HubSpot then blocks a new company with the same number, and imports can match on the number instead of the domain.
If you are moving from another CRM, clean up before the move. We describe the move in Migrating to HubSpot from another CRM.
Done when the company pairs are reviewed, the contacts at your largest customers are merged and the source that created the most duplicates is fixed.
Week 2: prune the fields
Where in HubSpot. The data quality overview (Data Management → Data Quality), which is in beta. On the Property insights tab HubSpot lists fields that look like duplicates of each other, fields with no data and fields that aren't used. The list updates daily, and from there you can hide or archive a field.
Go through the fields on companies, contacts and deals and make a decision for each one. This is what it can look like on the companies in an account that's a few years old:
| Field | Filled in | Decision |
|---|---|---|
| Industry | 4 in 10 companies | Keep as a dropdown. Fill in the rest with enrichment. |
| Industry 2, from a 2023 import | 2 in 10 | Move the values to Industry and archive. |
| Segment, free text: "SMB", "smb", "Small companies" | 3 in 10 | Turn it into a dropdown with three values. |
| Number of employees and Employees | 6 in 10 and 3 in 10 | Keep the first, move the values, archive the second. |
| Trade fair contact 2022 | under 1 in 100 | Archive. |
Merge fields that mean the same thing. Move the values from the old field to the one you're keeping, with a workflow or with an export and import. Then archive the old one. A field used in a workflow or form can't be archived until it's removed there. Also check reports and lists that use the field. An archived field can be restored for 90 days.
Dropdowns instead of free text. A free-text field ends up with several spellings of the same thing, and an agent filtering on segment only finds one of them. Turn fields with a limited set of answers into dropdowns. Formatting problems, such as capitalisation in names and email addresses, show up on the Formatting issues tab. There you accept suggestions one at a time, or all at once with Data Hub Professional or Enterprise. With Data Hub Professional or Enterprise, a rule can fix new issues automatically.
Required at the right stage. A field that's required when the record is created gets filled in with "x" when the rep doesn't know the answer yet. So make fields required at the deal stage where the answer is known. You set it per pipeline (Settings → Data Management → Objects → Deals → Pipelines). For each deal stage you choose the fields to fill in when a deal moves there, and tick that they're required. It applies when someone creates or moves a deal by hand, so check how it works for deals created by integrations and workflows. An example:
- Quote sent: amount, expected close date and who decides.
- Won: why the customer bought and who matters most at the customer. These are the fields that make the handover to customer success possible, see One customer, one place.
- Lost: the reason, as a dropdown.
When a company is created by hand, you can also require a few fields in the create form (Starter and up). Keep it short, for example name, domain and organisation number.
Validation rules stop the wrong format as the value is typed in, such as a phone number without a country code. They're available in every edition. Rules with regular expressions, which can check the format of an organisation number, need Professional or Enterprise.
Done when every field shown on the record has a reason to be there and the most important fields are required at the right deal stage. The number of fields with no data or no use should also have gone down in Property insights.
Week 3: write down the definitions
The fields say what's on the customer. The definitions say what it means. Take an example. Marketing counts everyone who has downloaded a guide as an MQL. Sales only counts the ones who have booked a demo. An agent asked to find qualified leads then gives different answers depending on whose number it reads.
Write a draft before the week starts. Then a meeting of just over an hour with the sales manager, the person responsible for marketing and the head of customer success is enough.
Lifecycle stages. An example:
| Lifecycle stage | What it means for us | Set by |
|---|---|---|
| Lead | Has left their details but hasn't been assessed | Form or import |
| MQL | Fits and is active: A1, A2 or B1 in lead scoring | Workflow |
| SQL | The rep has spoken to the person and wants to move forward | The rep |
| Opportunity | There's an open deal | Workflow when the deal is created |
| Customer | There's a won deal | Workflow when the deal is won |
We describe how the MQL threshold is calculated in Lead scoring in HubSpot.
Deal stages. Write each stage as something that has happened at the customer, so two reps put the same deal in the same stage.
| Deal stage | The deal is here when |
|---|---|
| Meeting booked | A first meeting is booked with someone who influences the purchase |
| Need confirmed | The customer has confirmed the problem and when they want to get started |
| Quote sent | The quote has gone to the person who decides |
| Negotiation | The customer has replied to the quote and you're discussing terms |
| Won | The contract is signed |
| Lost | The customer has said no or chosen someone else |
Who can change them. Only Super Admins can create and edit lifecycle stages in settings. With pipeline rules (Professional and Enterprise) you can stop deals from skipping stages or moving backwards, and limit who can edit deals in certain stages. With Enterprise you can also limit who can edit individual fields. HubSpot itself advises against treating that as a security control, since the value can still be changed through the API.
Put the definitions where the agents read. In Context Home you add things like your company profile, your products and how your teams work. Paste the definitions in there, so the assistant and the agents start from the same description as the teams.
Done when sales, marketing and customer success have signed off the page with lifecycle stages and deal stages. The stages in HubSpot use the same names as the page, and stages nobody uses are removed.
Week 4: owners and context
Owners. Every company, contact and open deal should have an owner who still works for you. Create saved views for companies without an owner, open deals without an owner and records owned by someone who has left. Then decide on a rule, for example:
- The company is owned by the rep or CSM who has the customer.
- Contacts get the same owner as the company, through a workflow.
- When someone leaves, their records are moved before the user is removed.
The owner matters for the agents too. Workflows and agents that create tasks often assign them to the record owner. If there's no owner, or the owner has left, the task goes nowhere.
Inboxes and calendars. Check that everyone who talks to customers has connected their inbox and calendar to HubSpot. Don't forget managers who handle their own customers. We cover the rest of the steps for gathering context, such as calls and meeting notes, in Context is king.
Context Home. Open Context Home and look at the score that shows how complete your context is and where the gaps are. Fill in the company profile and the products, and add the definitions from week 3. HubSpot also suggests changes when something in the context looks out of date.
One owner for data quality. Name one person, the data owner. In a team of five reps it can be the sales manager. With 30 reps it's often a separate role in the sales or RevOps team. Plan for a couple of hours a month for the checklist below. With Data Hub Professional or Enterprise, HubSpot can also email a data quality digest every Monday.
Done when the views for missing owners are empty and everyone who talks to customers has connected their inbox and calendar. The monthly review is also in the calendar as a recurring meeting.
Then: switch on one agent at a time
The agents live in Agent Hub. Start with one agent and one clearly defined task.
A good first test is the Data Agent, which answers questions about companies and contacts. Ask about a few customers where you already know the answer, for example whether the customer has an open ticket and when the contract ends. Compare the answer with what's on the company record. If they differ, there are still gaps in the data.
The Customer Agent answers from your knowledge base, so that's where the clean-up is needed for support. We describe how to set up customer support before you switch the agent on in Customer support in HubSpot. What the different agents do and what they cost is in AI in HubSpot.
The monthly checklist
This is the data owner's routine after the four weeks. Pull the numbers from day 0 every time and compare.
| What to check | Where in HubSpot | How often | Who |
|---|---|---|---|
| New duplicate pairs for companies and contacts | Manage duplicates | Every week | Data owner |
| Companies and open deals without an owner, or with an owner who has left | Saved views | Every week | Sales manager |
| Open deals without an amount or with a close date in the past | Saved view in deals | Every week | Sales manager |
| Fields with no data, unused fields and fields that look alike | Data quality, Property insights | Every month | Data owner |
| New fields someone has created | Settings → Properties | Every month | Data owner |
| Formatting issues | Data quality, Formatting issues | Every month | Data owner |
| Gaps in the context | Context Home | Every month | Data owner |
| Inbox and calendar connected | The new user's settings | With every new hire | Data owner |
| Contacts who haven't been active for a long time | A segment with your criteria | Every quarter | Marketing |
| The definitions still match how the teams work | The page from week 3 | Every quarter | Sales, marketing and customer success |
Old contacts: delete them or stop paying for them
Most accounts have thousands of contacts nobody has spoken to in years. They show up in searches and lists, for people and agents alike, and they're personal data you need a reason to keep.
GDPR. Personal data mustn't be kept longer than the purpose requires. So decide on a rule, for example that contacts with no activity for two years are deleted, except contacts at customers and contacts with an open deal. Have your lawyer or data protection officer confirm the period.
Automatic deletion. In HubSpot you can switch on automatic deletion of inactive contacts (Settings → Privacy & Consent). The default is 365 days, and you can set your own number of days. A contact counts as inactive when it hasn't visited the website, opened a marketing email or had any other recorded activity in that period. Deleted contacts stay in the recycle bin for 90 days and are then permanently deleted. Build a segment with the same criteria first and look at who would disappear. A contact at a customer who never opens your emails can look inactive. HubSpot also recommends thinking through how the deletion affects your reports.
Marketing contacts. If you have Marketing Hub, you pay for the contacts set as marketing contacts. Contacts you don't send marketing to can be set as non-marketing, and then you don't pay for them. Under Account & Billing there's a tool for cleaning up marketing contacts. The change applies from the next update date, and you don't drop to a lower tier until the licence renews.
Common mistakes
- Agent first, clean-up later. The agent's first weeks are when the teams decide whether they trust it. Wrong answers then are hard to repair.
- Twenty required fields when the record is created. Reps type "x", and the field becomes worse than empty.
- One big clean-up and then nothing. Without a data owner and a monthly routine, the account is soon back where it was.
If you want help
We do the clean-up as a defined RevOps project, which usually costs SEK 25,000–50,000. The monthly review of data and workflows is part of the Essential support plan, SEK 5,000 a month. All prices exclude VAT.
Common questions
How long does it take to clean up HubSpot? In an account with up to 30 reps and a couple of integrations it usually fits in four weeks. The data owner spends a few hours a week, and the teams meet once in week 3. If the account has years of imports and many integrations, the duplicates and the fields take longer. After that, a couple of hours a month is enough.
Which HubSpot licence do you need for the data quality tools? The data quality overview is available from Starter and is in beta. The duplicates tool needs Professional or Enterprise. Data Hub Professional or Enterprise is needed for bulk merging and automatic formatting fixes. The same goes for custom duplicate rules and rules for which record wins. The latter are in beta. Data Hub used to be called Operations Hub. Merging two records by hand works in every edition.
Can HubSpot clean up the data for us? Partly. HubSpot suggests formatting fixes and shows how similar two records are. HubSpot also has an automated health check of the data model with recommended fixes, which was improved in the autumn 2026 release. You still have to decide what an MQL is and which record should win. Let the tools suggest and a person approve, at least for the first few months.
Should we delete old contacts in HubSpot? Yes, the ones you have no reason to keep, following a rule you've decided in advance. Contacts you want to keep but not market to can be set as non-marketing contacts.
How often do we need to clean up HubSpot? Duplicates and deals without an owner every week, fields and Context Home every month, and the definitions every quarter.
Can we switch on the agents before everything is done? Yes, if the agent only works with the part of the data that's been cleaned up. An agent in the support chat needs an up-to-date knowledge base more than clean sales fields. Start where the data already holds up.
About the author
Viktor Olsson, founder and RevOps consultant
Viktor has worked with HubSpot for six years, with teams from 3 people to organisations with 30 sales reps across the Nordics. He also works in RevOps at an international SaaS company.